CMake API Reference#

nanobind’s CMake API simplifies the process of building python extension modules. This is needed because quite a few steps are involved: nanobind must build the module, a library component, link the two together, and add a different set of compilation and linker flags depending on the target platform.

If you prefer another build system, then you have the following options:

Nicholas Junge has created a Bazel interface to nanobind. Please report Bazel-specific issues there.
You could create a new build system from scratch that takes care of these steps. See this file for inspiration on how to do this on Linux. Note that you will be on your own if you choose to go this route—I unfortunately do not have the time to respond to GitHub tickets related to custom build systems.

The section on building extensions provided an introductory example of how to set up a basic build system via the nanobind_add_module() command, which is the high level build interface. The defaults chosen by this function are somewhat opinionated, however. For this reason, nanobind also provides an alternative low level interface that decomposes it into smaller steps.

A later part of this section explains how a Git submodule dependency can be avoided in exchange for a system-provided package.

Finally, the section ends with an explanation of the CMake convenience interface for stub generation.

High-level interface#

The high-level interface consists of just one CMake command:

nanobind_add_module#

Compile a nanobind extension module using the specified target name, optional flags, and source code files. Use it as follows:

nanobind_add_module(
  my_ext                   # Target name
  NB_STATIC STABLE_ABI LTO # Optional flags (see below)
  my_ext.h                 # Source code files below
  my_ext.cpp)

It supports the following optional parameters:

`STABLE_ABI`	Perform a stable ABI build, making it possible to use a compiled extension across Python minor versions. The flag is ignored on Python versions older than < 3.12.
`NB_STATIC`	Compile the core nanobind library as a static library. This simplifies redistribution but can increase the combined binary storage footprint when a project contains many Python extensions (this is the default).
`NB_SHARED`	The opposite of `NB_STATIC`: compile the core nanobind library as a shared library for use in projects that consist of multiple extensions.
`PROTECT_STACK`	Don’t remove stack smashing-related protections.
`LTO`	Perform link time optimization.
`NOMINSIZE`	Don’t perform optimizations to minimize binary size.
`NOSTRIP`	Don’t strip unneded symbols and debug information from the compiled extension when performing release builds.
`NB_DOMAIN <name>`	Restrict the inter-extension type visibility to a named subdomain. See the associated FAQ entry for details.
`MUSL_DYNAMIC_LIBCPP`	When cibuildwheel is used to produce musllinux wheels, don’t statically link against `libstdc++` and `libgcc` (which is an optimization that nanobind does by default in this specific case). If this explanation sounds confusing, then you can ignore it. See the detailed description below for more information on this step.

nanobind_add_module() performs the following steps to produce bindings.

It creates a CMake library via add_library(target_name MODULE ...) and enables the use of C++17 features during compilation.
It creates a CMake target for an internal library component required by nanobind (named nanobind-.. where .. depends on the compilation flags). This is only done once when compiling multiple extensions.

This library component can either be a static or shared library depending on whether the optional NB_STATIC or NB_SHARED parameter was provided to nanobind_add_module(). The default is a static build, which simplifies redistribution (only one shared library must be deployed).

When a project contains many Python extensions, a shared build is preferable to avoid unnecessary binary size overheads that arise from redundant copies of the nanobind-... component.
It links the newly created library against the nanobind-.. target.
It appends the library suffix (e.g., .cpython-39-darwin.so) based on information provided by CMake’s FindPython module.
When requested via the optional STABLE_ABI parameter, the build system will create a stable ABI extension module with a different suffix (e.g., .abi3.so).

Once compiled, a stable ABI extension can be reused across Python minor versions. In contrast, ordinary builds are only compatible across patch versions. This feature requires Python >= 3.12 and is ignored on older versions. Note that use of the stable ABI come at a small performance cost since nanobind can no longer access the internals of various data structures directly. If in doubt, benchmark your code to see if the cost is acceptable.
In non-debug modes, it compiles with size optimizations (i.e., -Os). This is generally the mode that you will want to use for C++/Python bindings. Switching to -O3 would enable further optimizations like vectorization, loop unrolling, etc., but these all increase compilation time and binary size with no real benefit for bindings.

If your project contains portions that benefit from -O3-level optimizations, then it’s better to run two separate compilation steps. An example is shown below:
```
# Compile project code with current optimization mode configured in CMake
add_library(example_lib STATIC source_1.cpp source_2.cpp)
# Need position independent code (-fPIC) to link into 'example_ext' below
set_target_properties(example_lib PROPERTIES POSITION_INDEPENDENT_CODE ON)

# Compile extension module with size optimization and add 'example_lib'
nanobind_add_module(example_ext common.h source_1.cpp source_2.cpp)
target_link_libraries(example_ext PRIVATE example_lib)
```
Size optimizations can be disabled by specifying the optional NOMINSIZE argument, though doing so is not recommended.
nanobind_add_module() also disables stack-smashing protections (i.e., it specifies -fno-stack-protector to Clang/GCC). Protecting against such vulnerabilities in a Python VM seems futile, and it adds non-negligible extra cost (+8% binary size in benchmarks). This behavior can be disabled by specifying the optional PROTECT_STACK flag. Either way, is not recommended that you use nanobind in a setting where it presents an attack surface.
It sets the default symbol visibility to hidden so that only functions and types specifically marked for export generate symbols in the resulting binary. This substantially reduces the size of the generated binary.
In release builds, it strips unreferenced functions and debug information names from the resulting binary. This can substantially reduce the size of the generated binary and can be disabled using the optional NOSTRIP argument.
Link-time optimization (LTO) is not active by default; benefits compared to pybind11 are relatively low, and this can make linking a build bottleneck. That said, the optional LTO argument can be specified to enable LTO in release builds.
nanobind’s CMake build system is often combined with cibuildwheel to automate the generation of wheels for many different platforms. One such platform called musllinux exists to create tiny self-contained binaries that are cheap to install in a container environment (Docker, etc.). An issue of the combination with nanobind is that musllinux doesn’t include the libstdc++ and libgcc libraries which nanobind depends on. cibuildwheel then has to ship those along in each wheel, which actually increases their size rather dramatically (by a factor of >5x for small projects). To avoid this, nanobind prefers to link against these libraries statically when it detects a cibuildwheel build targeting musllinux. Pass the MUSL_DYNAMIC_LIBCPP parameter to avoid this behavior.
If desired (via the optional NB_DOMAIN parameter), nanobind will restrict the visibility of symbols to a named subdomain to avoid conflicts between bindings. See the associated FAQ entry for details.

Low-level interface#

Instead of nanobind_add_module() nanobind also exposes a more fine-grained interface to the underlying operations. The following

nanobind_add_module(my_ext NB_SHARED LTO my_ext.cpp)

is equivalent to

# Build the core parts of nanobind once
nanobind_build_library(nanobind SHARED)

# Compile an extension library
add_library(my_ext MODULE my_ext.cpp)

# .. and link it against the nanobind parts
target_link_libraries(my_ext PRIVATE nanobind)

# .. enable size optimizations
nanobind_opt_size(my_ext)

# .. enable link time optimization
nanobind_lto(my_ext)

# .. set the default symbol visibility to 'hidden'
nanobind_set_visibility(my_ext)

# .. strip unneeded symbols and debug info from the binary (only active in release builds)
nanobind_strip(my_ext)

# .. disable the stack protector
nanobind_disable_stack_protector(my_ext)

# .. set the Python extension suffix
nanobind_extension(my_ext)

# .. set important compilation flags
nanobind_compile_options(my_ext)

# .. set important linker flags
nanobind_link_options(my_ext)

# Statically link against libstdc++/libgcc when targeting musllinux
nanobind_musl_static_libcpp(my_ext)

The various commands are described below:

nanobind_build_library#

Compile the core nanobind library. The function expects only the target name and uses a slightly unusual parameter passing policy: its behavior changes based on whether or not one the following substrings is detected in the target name:

`-static`	Perform a static library build (without this suffix, a shared build is used)
`-abi3`	Perform a stable ABI build targeting Python v3.12+.

# Normal shared library build
nanobind_build_library(nanobind)

# Static ABI3 build
nanobind_build_library(nanobind-static-abi3)

nanobind_opt_size#

This function enable size optimizations in Release, MinSizeRel, RelWithDebInfo builds. It expects a single target as argument, as in

nanobind_opt_size(my_target)

nanobind_set_visibility#

This function sets the default symbol visibility to hidden so that only functions and types specifically marked for export generate symbols in the resulting binary. It expects a single target as argument, as in

nanobind_trim(my_target)

This substantially reduces the size of the generated binary.

nanobind_strip#

This function strips unused and debug symbols in Release and MinSizeRel builds on Linux and macOS. It expects a single target as argument, as in

nanobind_strip(my_target)

nanobind_disable_stack_protector#

The stack protector affects the binary size of bindings negatively (+8% on Linux in benchmarks). Protecting from stack smashing in a Python VM seems in any case futile, so this function disables it for the specified target when performing a build with optimizations. Use it as follows:

nanobind_disable_stack_protector(my_target)

nanobind_extension#

This function assigns an extension name to the compiled binding, e.g., .cpython-311-darwin.so. Use it as follows:

nanobind_extension(my_target)

nanobind_extension_abi3#

This function assigns a stable ABI extension name to the compiled binding, e.g., .abi3.so. Use it as follows:

nanobind_extension_abi3(my_target)

nanobind_compile_options#

This function sets recommended compilation flags. Currently, it specifies /bigobj and /MP on MSVC builds, and it does nothing other platforms or compilers. Use it as follows:

nanobind_compile_options(my_target)

nanobind_link_options#

This function sets recommended linker flags. Currently, it controls link time handling of undefined symbols on Apple platforms related to Python C API calls, and it does nothing other platforms. Use it as follows:

nanobind_link_options(my_target)

nanobind_musl_static_libcpp#

This function passes the linker flags -static-libstdc++ and -static-libgcc to gcc when the environment variable AUDITWHEEL_PLAT contains the string musllinux, which indicates a cibuildwheel build targeting that platform.

The function expects a single target as argument, as in

nanobind_musl_static_libcpp(my_target)

Submodule dependencies#

nanobind includes a dependency (a fast hash map named tsl::robin_map) as a Git submodule. If you prefer to use another (e.g., system-provided) version of this dependency, set the NB_USE_SUBMODULE_DEPS variable before importing nanobind into CMake. In this case, nanobind’s CMake scripts will internally invoke find_dependency(tsl-robin-map) to locate the associated header files.

Stub generation#

Nanobind’s CMake tooling includes a convenience command to interface with the stubgen program explained in the section on stub generation.

nanobind_add_stub#

Import the specified module (MODULE parameter), generate a stub, and write it to the specified file (OUTPUT parameter). Here is an example use:

nanobind_add_stub(
    my_ext_stub
    MODULE my_ext
    OUTPUT my_ext.pyi
    PYTHON_PATH $<TARGET_FILE_DIR:my_ext>
    DEPENDS my_ext
)

The target name (my_ext_stub in this example) must be unique but has no other significance.

stubgen will add all paths specified as part of the PYTHON_PATH block and then execute import my_ext in a Python session. If the extension is not importable, this will cause stub generation to fail.

This command supports the following parameters:

`INSTALL_TIME`	By default, stub generation takes place at build time following generation of all dependencies (see `DEPENDS`). When this parameter is specified, stub generation is instead postponed to the installation phase.
`MODULE`	Specifies the name of the module that should be imported. Mandatory.
`OUTPUT`	Specifies the name of the stub file that should be written. The path is relative to `CMAKE_CURRENT_BINARY_DIR` for build-time stub generation and relative to `CMAKE_INSTALL_PREFIX` for install-time stub generation. Mandatory.
`PYTHON_PATH`	List of search paths that should be considered when importing the module. The paths are relative to `CMAKE_CURRENT_BINARY_DIR` for build-time stub generation and relative to `CMAKE_INSTALL_PREFIX` for install-time stub generation. The current directory (`"."`) is always included and does not need to be specified. The parameter may contain CMake generator expressions when `nanobind_add_stub()` is used for build-time stub generation. Otherwise, generator expressions should not be used. Optional.
`DEPENDS`	Any targets listed here will be marked as a dependencies. This should generally be used to list the target names of one or more prior `nanobind_add_module()` declarations. Note that this parameter tracks build-time dependencies and does not need to be specified when stub generation occurs at install time (see `INSTALL_TIME`). Optional.
`VERBOSE`	Show status messages generated by `stubgen`.
`EXCLUDE_DOCSTRINGS`	Generate a stub containing only typed signatures without docstrings.
`INCLUDE_PRIVATE`	Also include private members, whose names begin or end with a single underscore.
`MARKER_FILE`	Typed extensions normally identify themselves via the presence of an empty file named `py.typed` in each module directory. When this parameter is specified, `nanobind_add_stub()` will automatically generate such an empty file as well.
`PATCH_FILE`	Specify a patch file used to replace declarations in the stub. The syntax is described in the section on stub generation.
`COMPONENT`	Specify a component when `INSTALL_TIME` stub generation is used. This is analogous to `install(..., COMPONENT [name])` in other install targets.
`EXCLUDE_FROM_ALL`	If specified, the file is only installed as part of a component-specific installation when `INSTALL_TIME` stub generation is used. This is analogous to `install(..., EXCLUDE_FROM_ALL)` in other install targets.